Identifying Automatically Generated Headlines using Transformers

Code for NLP4IF 2021 workshop paper.

How to replicate results?

Step 1: Download Data.

• Link to generated headlines data (attack).
• Link to generated headlines data (defense).
• For the real headlines, go to Million Headlines, or contact me for the final dataset (due to copyright concerns I cannot share the data myself publicly).

Step 2: Pretrain two LMs using the HuggingFace library. Pretrain one LM on the "attack" data and another on the "defense". Sample bash script at gpt2_pretraining.sh.

Step 3: Generate headlines for attack and defense (if you haven't downloaded the ones from Step 1).

Step 4: Merge the real/generated headlines for attack and defense.

Step 5: Finetune your classifiers (using the example scripts provided here, ie. *_classify.py).